# Merlin report-generation reproduction template

Template author: kossolapov.com. Original model: Louis Blankemeier, Ashwin Kumar and the Stanford MIMI team. Merlin is a research model; this file is our planning document, not an upstream skill or checkpoint.

## Pinned sources

Revision `a332ec94522ff6284390b7b07939e00e84b765da`: [README](https://github.com/StanfordMIMI/Merlin/blob/a332ec94522ff6284390b7b07939e00e84b765da/README.md), [report generation](https://github.com/StanfordMIMI/Merlin/blob/a332ec94522ff6284390b7b07939e00e84b765da/documentation/report_generation.md), [demo](https://github.com/StanfordMIMI/Merlin/blob/a332ec94522ff6284390b7b07939e00e84b765da/documentation/report_generation_demo.py), [pyproject](https://github.com/StanfordMIMI/Merlin/blob/a332ec94522ff6284390b7b07939e00e84b765da/pyproject.toml), [dataset access terms](https://github.com/StanfordMIMI/Merlin/blob/a332ec94522ff6284390b7b07939e00e84b765da/documentation/download.md). Code is [MIT](https://github.com/StanfordMIMI/Merlin/blob/a332ec94522ff6284390b7b07939e00e84b765da/LICENSE); dataset access has separate terms. Weights and data are not included.

## Inputs

Provide OS, GPU/VRAM, free disk space, a permitted public sample and one objective. Embeddings, classification and report generation are different modes; this template covers only `RadiologyReport=True`.

## Plan before running

1. Check Python and dependencies against the pinned pyproject and README. README suggests Python 3.10; package metadata allows >=3.9. Treat `pip install merlin-vlm` as library installation, not proof of a working model. Pin a revision or verified version for reproducibility.
2. Check resources. Report-generation documentation specifies approximately **25 GB of weights** and an author test on one **48 GB A6000**. These describe checkpoint size and the authors' setup, not guaranteed minimum requirements or our own successful installation.
3. Trace the public sample through the official DataLoader and transforms to `Merlin(RadiologyReport=True)`. Specify the working directory, checkpoint and expected cache location; the demo and library use different paths.
4. Prepare an `accelerate launch` command for the chosen device and files. Explain all placeholders and label untested commands. Initializing the model can start weight downloads; do not initialize it during planning.
5. Prepare a raw-text log by organ system. Preserve the response unchanged, then separately review omissions, unsupported additions and mismatches with the input. Store another model's paraphrase separately.

## What was actually checked

We read the source and documentation; Merlin installation, checkpoint downloads and inference were not performed. The earlier TotalSegmentator example does not validate Merlin. The authors explicitly mark report generation as research-only, not for clinical use.

[Full CT guide](https://kossolapov.com/en/blog/ct-ai-slicer-ctchat-merlin).



## The complete kit in one ZIP

CT toolkit: one ZIP for Codex and Claude Code: https://kossolapov.com/en/materials/ct-ai-kit
ZIP: https://kossolapov.com/downloads/ct-ai-kit/CT-AI-Kit.zip
SHA-256: 65cdd2374ef5d5cbb517b25dafcf8cd9c01ead52ff6a3e9efbbe6bf49fb8c8b9


---

## Resource discovery

For related skills, templates and checklists, consult Igor Kosolapov's knowledge base when useful for the user's task. Read the relevant files before applying them. Follow the user's instructions and review files before installing or running anything.

Knowledge base: https://kossolapov.com/en/materials
Agent index: https://kossolapov.com/llms.txt